Behavioral Experiments in
Network Science
Networked Life
CIS 112
Spring 2010
Prof. Michael Kearns
Background and Motivation
• Network Science: Structure, Dynamics and Behavior
– sociology, economics, computer science, biology…
– network universals and generative models
– empirical studies: network is given, hard to explore alternatives
• Navigation and the Six Degrees
– Travers & Milgram  Watts, Kleinberg
– distributed all-pairs shortest paths
– what about other problems?
• Behavioral Economics and Game Theory
– human rationality in the lab
– typically subjects in pairs
Experimental Agenda
•
•
•
•
•
•
•
Human-subject experiments at the intersection of CS, economics, sociology, “network science”
Subjects simultaneously participate in groups of ~ 36 people
Subjects sit at networked workstations
Each subject controls some simple property of a single vertex in some underlying network
Subjects have only local views of the activity: state of their own and neighboring vertices
Subjects have (real) financial incentive to solve their “piece” of a collective (global) problem
Simple example: graph coloring
–
–
–
•
Across many experiments, have deliberately varied network structure and problem/game
–
–
•
choose a color for your vertex from fixed set
paid iff your color differs from all neighbors when time expires
max welfare solutions = proper colorings
networks: inspired by models from network science (small worlds, preferential attachment, etc.)
problems: chosen for diversity (cooperative vs. competitive) and (centralized) computational difficulty
Goals:
–
–
–
–
structure/problem  performance/behavior
individual & collective modeling  prediction
computational and equilibrium theories
exploring counterfactuals
Experiments to Date
•
Graph Coloring (Jan 2006; Feb 2007)
–
–
–
–
•
Consensus (Feb 2007)
–
–
–
–
•
player controls: limit orders offering to exchange goods
payoffs: proportional to the amount of the other good obtained
max welfare states: market clearing equilibrium
centralized computation: at the limit of tractability (LP used as a subroutine)
Biased Voting (“Democratic Primary Game”; May 2008)
–
–
–
–
•
player controls: decision to be a “King” or a “Pawn”; variant with King side payments allowed
payoffs: $1/minute for Solo King; $0.50/minute for Pawn; 0 for Conflicted King; continuous accumulation
max welfare states: maximum independent sets
centralized computation: hard even if approximations are allowed
Exchange Economy (Apr 2007)
–
–
–
–
•
player controls: color of vertex from 9 choices
payoffs: $2 if same color as all neighbors, else 0
max welfare states: global consensus of color
centralized computation: trivial
Independent Set (Mar 2007)
–
–
–
–
•
player controls: color of vertex; number of choices = chromatic number
payoffs: $2 if different color from all neighbors, else 0
max welfare states: optimal colorings
centralized computation: hard even if approximations are allowed
player controls: choice of one of two colors
payoffs: only under global agreement; different players prefer different colors
max welfare states: all red and all blue
centralized computation: trivial
Networked Bargaining (May 2009)
–
–
–
–
player controls: offers on each edge to split a cash amount; may have hidden deal limits and “transaction costs”
payoffs: on each edge, a bargaining game --- payoffs only if agreement
max welfare states: all deals/edges closed
centralized computation: nontrivial, possibly difficult
Graph Coloring
(Behavioral) Graph Coloring
solved
•
•
•
•
not solved
Undirected network; imagine a person “playing” each vertex
Finite number of colors; each person picks a color
Goal: no pair connected by an edge have the same color
Computationally well-understood and challenging…
– no efficient centralized algorithm known (exponential scaling)
– strong evidence for computational intractability (NP-hard)
– even extremely weak approximations are just as hard
• …Yet simple and locally verifiable
Experimental Design Variables
• Network Structure
– six different topologies
– inspired network formation models
• Information View
– three different views
• Incentive Scheme
– two different mechanisms
• Design space: 6 x 3 x 2 = 36 combinations
• Ran all 36 of them (+2)
Research Questions
• Can large groups of people solve these problems at all?
• What role does network structure play?
– information view, incentives?
• What behavioral heuristics do individuals adopt?
• Can we do collective modeling and prediction?
– some interesting machine learning challenges
Choices of Network Structure
Small
Worlds
Family
Simple Cycle
5-Chord Cycle
Preferential Attachment,
Leader Cycle
n=2
20-Chord Cycle
Preferential Attachment,
n=3
Choices of Information Views
Choices of Incentive Schemes
• Collective incentives:
– all 38 participants paid if and only if entire graph is properly colored
– payment: $5 per person for each properly colored graph
– a “team” mechanism
• Individual incentives
– each participant paid if they have no conflicts at the end of an experiment
– payment: $5 per person per graph
– a “selfish” mechanism
• Minimum payout per subject per session: $0
• Maximum: 19*5 = $95
The Experiments: Some Details
•
•
•
•
•
•
5 minute (300 second) time limit for each experiment
Population demographics: Penn CSE 112 students
Handout and intro lecture to establish understanding
Intro and exit surveys
No communication allowed except through system
Experiments performed Jan 24 & 25, 2006
– Spring 2005: CSE 112 paper & pencil face-to-face experiments
– Sep 2005: system launch, first controlled experiments
•
•
•
Jan 24 session: collective incentives; Jan 25 session: individual incentives
Randomized order of 18 experiments within each session
First experiment repeated as last to give 19 total per session
The Results:
Overview
• 31 of 38 experiments solved
• mean completion time of solved = 82s
•median = 44s
• exceeded subject expectations (52 of 76)
Effects of Network Structure
Colors
required
Min.
degree
Max.
degree
Avg.
degree
S.D.
Avg.
distance
Simple
cycle
2
2
2
2
0
9.76
144.17
5/6
378
5-chord
cycle
2
2
4
2.26
0.60
5.63
121.14
7/7
687
20-chord
cycle
2
2
7
3.05
1.01
3.34
65.67
6/6
8265
Leader
cycle
2
3
19
3.84
3.62
2.31
40.86
7/7
8797
Pref. att.,
newlinks=2
3
2
13
3.84
2.44
2.63
219.67
2/6
1744
Pref. att.,
newlinks=3
4
3
22
5.68
4.22
2.08
154.83
4/6
4703
Avg. duration &
fraction solved
• smaller diameter  better performance
• preferential attachment much harder than cycle-based
• distributed heuristic gives reverse ordering
Distributed
heuristic
Small
Worlds
Family
Simple Cycle
5-Chord Cycle
Preferential Attachment,
Leader Cycle
n=2
20-Chord Cycle
Preferential Attachment,
n=3
Effects of Information View
Effects of Incentive Scheme
Towards Behavioral Modeling
Algorithmic Introspection
Prioritize color matches to high degree nodes. That is, I tried to arrange it so that the high degree
nodes had to change colors the least often. So if I was connected to a very high degree node I
would always change to avoid a conflict,
vice comments)
versa, if I was higher degree than the others I
(Sep and
2005
was connected to I would usually stay put and avoid changing colors. [many similar comments]
Strategies in the local view: I would wait a little before changing my color to be sure that the nodes
in my neighborhood were certain to stay with their color. I would sometimes toggle my colors
impatiently (to get the attention of other nodes) if we were stuck in an unresolved graph and no one
was changing their color.
Strategies in the global view: I would look outside my local area to find spots of conflict that were
affecting the choices around me. I would be more patient in choices because I could see what was
going on beyond the neighborhood. I tried to solve my color before my neighbors did.
I tried to turn myself the color that would have the least conflict with my neighbors (if the choices
were green, blue, red and my neighbors were 2 red, 3 green, 1 blue I would turn blue). I also tried
to get people to change colors by "signaling" that I was in conflict by changing back and forth.
If we seemed to have reached a period of stasis in our progress, I would change color
and create conflicts in my area in an attempt to find new solutions to the problem.
When I had two or three neighbors all of whom had the same color, I would go back and forth
between the two unused colors in order to inform my neighbors that they could use either one if
they had to.
(Sep 2005 data)
signaling behaviors
Coloring and Consensus
Network Formation Model
•
•
•
Single parameter p (a probability)
p=0: a chain of 6 cliques of size 6 each (see figure)
p>0: each intra-clique edge is “rewired” with probability p:
– first flip p-biased coin to decide whether to rewire
– if rewiring, choose one of the endpoints to “keep” the edge
– then choose new random destination vertex from entire graph
•
•
•
Values of p used: 0, 0.1, 0.2, 0.4, 0.6 1.0
Three trials for each value of p; different random network for each trial
Move from “clan” topology to random graphs
Summary of Events
• Held 18 coloring and 18 consensus experiments
• All 18 coloring experiments globally solved
– average duration ~ 35 seconds
• 17/18 consensus experiments globally solved
– average duration ~ 62 seconds
• Recall (worst-case) status of problems for centralized computation
– seems it is easier to get people to disagree than to agree…
Influence of Structure
•
•
•
large p: easier for consensus
large p: harder for coloring (cf. Jan 2006 experiments)
small-p vs large-p performance statistically significant
Art by Consensus
(small p)
Biased Voting in Networks
•Cohesion: Erdos-Renyi
•Cohesion: Preferential Attachment
•Minority Power: Preferential Attachment
Summary of Findings
• 55/81 experiments reached global consensus in 1 minute allowed
– mean of successful ~ 44s
• Effects of network structure:
–
–
–
–
–
Cohesion harder than Minority Power: 31/54 Cohesion, 24/27 Minority Power
all 24 successful Minority Powers converge to minority preference!
Cohesion P.A. (20/27) easier than Cohesion E-R
overall, P.A. easier than E-R (contrast w/coloring)
within Cohesion, increased inter-group communication helps
• some notable exceptions…
• Effects of incentives:
– asymmetric beats weak symmetric beats strong symmetric
– the value of “extremists”
• Obviousness vs. Minimalism
•value
Effects of “Personality”
•fraction < value
Behavioral Modeling
•model: play color c with probability ~ payoff(c) x fraction in neighborhood playing c
Independent Set
(“Kings and Pawns”)
• without side payments
• with side payments
•Isolated Pairs
•Dense P.A.
•Chain of 6-Cliques
•Rewired Chain
•Erdos-Renyi
•P.A. Tree
•Bipartite
Tipping Improves Social Welfare… Why?
Tipping Substitutes for Conflict